DeepSeek Intensifies AI Race with New Model Tailored for Chinese Chips
DeepSeek AI has released its latest large language model, DeepSeek-V3.1, escalating the global artificial intelligence race with a new architecture designed to run on next-generation domestic Chinese chips and capabilities that challenge leading Western models.
The new model introduces a novel hybrid inference architecture and demonstrates significantly enhanced performance in complex "agent" tasks such as programming and multi-step search. Early third-party benchmarks show DeepSeek-V3.1 outperforming prominent models like Anthropic’s Claude 3 Opus on coding tests, while maintaining a substantial cost advantage, signaling a new competitive threat in the high-stakes AI market.
In a clear move to accelerate its commercialization strategy, DeepSeek announced it will implement a new API pricing structure and eliminate its long-standing nighttime discounts starting September 6, 2025. This move follows a service capacity expansion and aims to capitalize on the model's enhanced capabilities and growing developer interest.
Critically, the model's design incorporates the UE8M0 FP8 parameter precision format, a technical specification explicitly tailored for upcoming domestic semiconductor hardware. This development underscores a broader strategic push within China's tech sector to build a self-reliant AI ecosystem, from foundational models to the underlying silicon.
Hybrid Architecture Drives Agent Performance
At the core of DeepSeek-V3.1 is its hybrid inference architecture, which allows the model to dynamically switch between a "thinking" mode for complex reasoning and a "non-thinking" mode for straightforward tasks within a single framework. The company claims this approach enhances efficiency, with the new "Think" mode reducing token consumption by 20% to 50% compared to its predecessor while achieving similar performance.
The update places a strong emphasis on strengthening the model’s agent capabilities. According to company data, V3.1 shows marked improvement in software engineering tasks (SWE-bench), command-line operations (Terminal-Bench), and complex search queries requiring multi-step reasoning (browsecomp). These enhancements are designed to make the model more attractive for commercial applications involving automation and complex information processing.
Benchmark Data Shows Competitive Edge
Independent testing has corroborated DeepSeek's claims of a significant leap in performance, particularly in programming. On the widely-recognized Aider multi-language coding benchmark, V3.1 achieved a score of 71.6%, surpassing several established models, including Claude 3 Opus.
The benchmark data also highlights the model's cost-efficiency. Completing a standard programming task on the Aider test cost just $1.01 with DeepSeek-V3.1, a fraction of the cost associated with some proprietary models. This combination of high performance and low operational cost is a key part of its value proposition for developers and enterprise users, who are increasingly sensitive to the high price of advanced AI.
API Upgrades and Open-Source Strategy
Continuing its commitment to the open-source community, DeepSeek has released the 685-billion-parameter base and post-trained versions of V3.1 on platforms like Hugging Face. The company noted the new model has been trained on an additional 840 billion tokens and requires developers to use an updated tokenizer and chat template.
Simultaneously, the company has upgraded its API services. The context window has been expanded to 128K, and developer tools have been enhanced with features like a "strict" mode for Function Calling to ensure reliable output formatting. To lower the barrier for adoption, the API now also supports the Anthropic API format, simplifying the migration process for developers already using frameworks built for Claude models.
A Clearer Path to Commercialization
Beyond the technical upgrades, DeepSeek is taking decisive steps toward monetization. The company will overhaul its API pricing effective September 6, 2025. The new structure will charge for input at RMB 0.5 per million tokens (about $0.07) for cache hits and RMB 4 for cache misses. Output will be priced at RMB 12 per million tokens. The existing nighttime discount policy will be discontinued a day prior.
DeepSeek stated that the pricing adjustment is intended to better meet surging user demand following a recent expansion of its API service resources. The move is widely seen as a pivotal step for the company as it transitions from a phase of rapid model development and user acquisition to building a sustainable commercial operation.